A hybrid neural network and Minimax algorithm for zero-sum games

摘要

This paper describes a hybrid approach to creating computer players for zero-sum games. This novel approach consists of a modified Minimax algorithm and a neural network. The standard Minimax algorithm requires too much processing time to be useful in all but the simplest games. Utilizing a neural network to limit the size of the game tree searched by Minimax greatly reduces the processing time required. A case study of Tic-Tac-Toe on larger boards was implemented to validate the new approach. Experimental evidence is presented that indicates that the suggested algorithm yields an effective player even though the search space is substantial.